How Can I Automate Data Extraction From Complex Documents?
Can you extract the full value of data and information from complex documents? Find out how to make AI technologies work for you.
Join the DZone community and get the full member experience.Join For Free
Business processes fed by complex documents are a bear.
NO!! Not that type of bear...This type of bear!
In a place where complicated can slow things down to a crawl, complex documents suck the life out of productivity.
Sure, you might have an OCR system in place that processes your documents.
And OCR is a good technology...for structured documents. But what about those complex, unstructured docs?
Or heck, maybe you're still manually processing your documents. Good ol’ human effort is a tried and true way to key a document into the system that runs your business process. A human can even find the right data in a sea of complex data. Eventually.
But, humans are slow, error-prone, inconsistent, and expensive. (And, in some cases, perhaps not so excellent after all!)
Then, there are all the challenges.
- Can have multiple formats
- Can’t be forced into a template
- Maybe free-flowing
- Might have tables...or worse! Nested tables!
- Could feature images
- Might include hand-writing…or worse! Messy handwriting!
- [FILL IN YOUR OWN FAVORITE EXTRACTION PAIN HERE!]
The worst part? OCR systems definitely hit a wall when documents get too complex.
So much for automation, right?
(Alas, fine reader...there is hope.)
What Is a Document-Centric Workflow?
In its simplest form, a document-centric workflow is one that executes a business process. In almost all cases, documents feed the process, which includes capturing content, extracting information from the content, and taking some action based on that information.
For example, here’s a document feed process that probably sounds familiar....
I submit a healthcare expense to my health insurance to get reimbursed. I have to:
- Copy the receipt
- Print out forms
- Fill out the forms
- Get an envelope and stamp
- Figure out the address
- Mail it
And, that’s just my end.
In process-centric workflow use cases, content contains data and information that’s contextually relevant to the process and the business.
The content we’re all using has value trapped in it...value that’s tough to release.
Documents can be classified into various forms and types. Documents can be images, text, numbers, videos, or a mix of types.
Classification can be based on any number of things, including:
- Annual Reports
- Bank statements
- ACORD forms
- Handwritten forms
- Utility bills
- Electrical panel
- And a whole lot more!
Information trapped in the documents can be extracted using a manual process, OCR, or some other technology. When deciding which of these to use, it’s important to know if we can extract all the information in the doc and how accurate that information is.
Then, extracted data and information are fed into a process. Think Mortgage Processing, Itinerary Processing, Loan Processing, Claims Processing, RFP Response Processing, Financial Compliance, Auditing, Expense Management, Invoice Processing, and so on.
You likely have been executing processes that require data extraction for some time. If you’re like most, you’ve run into roadblocks. And because of those roadblocks, your automation plans are stuck.
The culprit? It’s probably complex data.
How to Tell if Your Complex Data Blocks Your Automation Goals?
There’s a good reason for more process automation where possible. 10x+ improvement in efficiency, productivity, and/or cost savings sounds incredible, right?!
If your goal is to automate more of these document-fed processes that now require humans for data entry...or the ones that OCR proves it can’t handle, how do you diagnose the problem so you can meet your goals?
And, how do you know when complex data is creating a process bottleneck?
The complexity of your data likely indicates the level of difficulty you’ll face when trying to extract the data and draw insights from it.
What are some factors that make documents complex to process?
- Content is free-flowing
- The document is unstructured
- It contains handwriting
- It is made up of multiple document types
- Formats change in the same doc
- Fonts change in the same doc
- The document has complex tables
- Tables are in different locations
- There is missing information
- Pictures and images are present
These are document types where OCR fails, and manual processing becomes overly complicated.
What Is the Business Result of Complex Documents?
When you have complex documents that cannot be automated, your business suffers.
What does it look like?
- High operational costs
- Low process efficiency
- Long process completion times
- Extraction accuracy that’s too low to be useful
I think these customers nailed it when they said...
“As a financial company, our employees spend a lot of time rewriting invoices.”
“We want to extract all the info from docs, so we can automate more processes and use all the info to build insights. But our analysts use only 10-20% of the data in the documents because we cannot extract the rest.”
Solutions for Complex Data Processing
The industry has evolved from OCR to solutions that use multiple AI technologies to address the bottlenecks. These solutions are categorized by:
- The old-school approach: OCR
- The modern approach: Various names, including:
- Intelligent Data Processing
- Intelligent Data Capture
- Machine Learning OCR
- Cognitive Capture
- AI OCR
- AI RPA
Elsewhere you will read how AI technology is being applied to solve unstructured data problems. Be cautious here; AI has become a buzzword some vendors deploy to cloud the waters when it comes to describing how AI plays in their solutions.
For now, the key point is this:
Intelligent Document Processing (IDP) can extract virtually all the information, understand the data, and create additional value from complex documents.
The Three Most Common Problems of Complex Docs
Infrrd has worked hand-in-hand with hundreds of enterprises and companies to solve complex data problems. We have lots of stories to share. For now, let's review the top three use cases we encounter most often.
Problem 1. Data Extraction From Annual Reports
A financial services company provides business loans.
The bank both originates and services the loan. The firms they lend to must submit financial reports so the bank can ensure financial soundness and compliance.
Pretty straightforward, right? So what’s the problem?
Financial reports (annual reports in this case) have no universal standard; They usually come in different formats, have non-standard taxonomies, and can vary year-to-year. These reports include graphics, charts, and tables, which are also inconsistent.
The complexity of these documents requires manual processing because OCR can’t handle the doc with so little structure. What’s worse? This manual process is always more costly, slower, and inconsistent. Even the smallest error can call into question the bank’s entire financial evaluation.
But, without the information trapped in these documents, the bank cannot determine how well the firms in its loan portfolio are doing and why. And when the information isn’t delivered on a timely basis? That’s when the bank introduces unnecessary operating risks into its system.
We worked with this bank to extract data from their complex documents. The bank now uses our Intelligent Data Processing solution, which applies a multi-layered sequence of AI models. The result? This bank no longer has an annual report processing problem.
Problem 2. Data Extraction From Panel Drawings
A panel drawing is an image that describes the layout and components of a control panel, a distribution panel, or an electrical panel.
The sample below shows there are part numbers and specifications for the components as well.
So how do you extract usable data from these panels? Are they too complex to do so?
A supplier receives an RFP package from a builder which includes documents and panel drawings. The supplier needs to read the drawings, build a quote, and send it to the builder. If the supplier has the best quote, they win the business.
But when the RFP package (docs and lots of panel drawings) is processed manually, it takes weeks to build a quote.
Could automated data extraction be used on these panel drawings?
We found out from working with this supplier that they tried OCR...and failed.
OCR cannot process panel drawings because it fails to:
- Identify line style and thickness
- Understand text orientation (top, bottom, side of drawing)
- Differentiate symbols from numbers and letters
The supplier—after partnering with us—learned how to use an AI-native information extraction platform to address the unique challenges of even the most complex panel drawings. As a result, the supplier automated their RFP process. Today they respond to the builders they serve 20x faster and with higher accuracy.
Contrary to popular opinion, YES. You can automate data extraction from panel drawings.
Problem 3. Data Extraction From Tables
Tables are everywhere. You’ll find them in annual reports, financial statements, invoices, bills, receipts, and management reports.
Tables help structure information, so we humans can more easily understand it.
And...tables really are everywhere. Odds are they’re in the very documents that contain the information you want to extract!
The biggest challenge with tables shows its face as complexity increases. Here’s what that looks like:
- Tables don’t appear in the same place in reports
- Fonts vary in the same table
- There are numbers and letters in the table
- Tables show up with and without borders
- You find tables within tables (nested tables)
- Tables go on for tens—or even hundreds—of pages
Manual processing of tables may work in the case of a simple table with limited rows and columns. But when tables extend across many pages, anyone reading the data can make mistakes.
As you’d imagine, OCR is challenged by tables, too. When a table is without borders—like below—an OCR fails to identify the information as a table...and, of course, the table type.
OCR also fails when it has to identify if an entry is a zero or an “O.”
We and our clients have been successfully extracting data from tables for a long time. It takes a different mindset and an approach completely different from OCR to consistently get it right.
Dropping Knowledge Bombs on Information Extraction
In this blog, you’ve learned some basics of data extraction from complex documents.
Remember the three challenging use cases (Annual Reports, Panels and Tables)? Most people who experience these throw up their hands in frustration…and walk away. They never tap into the true value that’s trapped in their documents!
Can you extract the full value of data and information from complex documents?
YES. YOU. CAN.
Explore our blog posts to learn how to solve each of these unstructured data problems.
We’ll discuss it all in more detail.
And, you’ll see just how to make AI technologies work for you.
You’ll become the complex data extraction maestro of your organization. And the automation angels will sing your name in unison.
But, look out! There will be quizzes, and you’ll have to put on that thinking cap!
Until then, ponder this: What else could we achieve if we could extract all the data and information from all of our complex documents?
The answer to that will likely astound you.
Until next time...
Published at DZone with permission of Mark Clark. See the original article here.
Opinions expressed by DZone contributors are their own.
10 Traits That Separate the Best Devs From the Crowd
Strategies for Reducing Total Cost of Ownership (TCO) For Integration Solutions
Integrating AWS With Salesforce Using Terraform
How To Use Pandas and Matplotlib To Perform EDA In Python